Visual Detection Method Of Substation Equipment In Haze Weather And Occlusion Scene

The substation is a very important part in the power system,and the real-time status of the substation equipment is also related to the safety of the substation.The traditional inspection method has shortcomings such as low efficiency and high labor intensity,which cannot meet the needs of modern production.Robot inspection instead of manual inspection has become an inevitable trend of modern substation maintenance.During the inspection process of the robot,the haze weather makes the visibility of the substation equipment low and the target unclear.In addition,the parameters of the commonly used target detection algorithms are too large,and it is difficult to meet the real-time requirements when deployed on inspection robots.To this end,the paper mainly focuses on the identification of substation equipment in haze weather and occlusion scenarios,and studies a lightweight target detection algorithm to deploy on inspection robots to achieve real-time detection.A foggy substation image dehazing algorithm based on improved CycleGAN is proposed.Compared with the classic dehazing algorithm,it does not require paired data sets for learning and training,which reduces the difficulty of data set production,and then introduces color loss into the original CycleGAN network.,to solve the problem of color distortion after dehazing,and finally introduce perceptual loss to make the details of the dehazed picture more similar to the original picture.In order to realize real-time detection of robots,the identification method of lightweight network is adopted.First,MobileNetv3 is used to improve the backbone feature extraction network of YOLOv4,and depth wise separable convolution is used to improve and strengthen the 3 X 3 ordinary convolution in the feature extraction network PANet.Lightweight,improves the speed of detection,and then introduces the SKNet attention mechanism into the detection head to improve the accuracy of target recognition on the basis of light weight,and deploys the algorithm in the robot host computer Qt.Aiming at the problem that the state recognition accuracy of the isolation knife gate is low in the occlusion scene,the Repulsion loss function that improves the exclusion term is introduced into the loss function of the lightweight YOLOv4 algorithm,and the Soft-DIoU-NMS soft non-maximum suppression is improved.The algorithm ensures that the state of the isolation knife gate can be accurately distinguished under the condition of occlusion.Through field experiments,the dehazing algorithm of the improved CycleGAN can be used to complete the dehazing processing of the real scene of the substation.The SSIM value of the image after dehazing reaches more than 0.8,the image distortion after dehazing is less,and the image details are closer to the original image;the improved The light-weight substation power equipment target detection model greatly improves the detection speed on the premise of ensuring the detection accuracy.Compared with YOLOv4,the number of model parameters is reduced by 80.21%.The precision mAP value reaches 89.4%,which meets the requirements of real-time inspection by the inspection robot;in the case of occlusion,the state recognition AP value of the isolation knife gate reaches 82.16%,which is about 8.47%higher than the original YOLOv4.There are fewer inspections,and its status can be accurately distinguished even under occlusion.